Latest / Pediatrics Now: Cases Updates and Discussions for the Busy Pediatric Practitioner / AI & Medicine: the Chief Medical Officer for Microsoft on the Next Revolution in Clinical Care

Transcript

AI & Medicine: the Chief Medical Officer for Microsoft on the Next Revolution in Clinical Care · Pediatrics Now: Cases Updates and Discussions for the Busy Pediatric Practitioner

0:00
I'm Holly Wehment, and this is Pediatrics Now, cases, updates,
0:04
and discussions for the busy pediatric practitioner.
0:07
What is the future of human health when it comes to AI and medicine?
0:12
I'm at a fascinating conference in downtown San Antonio, brought to you by UT San Antonio.
0:18
The chief medical officer for Microsoft, Eric Horvitz, gives a fascinating talk. Let's listen in.
0:26
This world is evolving so quickly in the domains that you're operating in right
0:31
now, the domain of computing and cyber and AI, especially, and data science.
0:39
All of these things have converged into something that's going at hyperspeed these days.
0:44
And the best way to create the future, carve out the future,
0:49
is when you bring ideas together.
0:50
And an ideation summit, a density of excellence of ideas and bringing them all
0:56
together is really powerful. And that's why I want this to continue and go on every year.
1:01
So to my colleagues who organize this, thank you. And to all of you who are
1:05
attending, thank you for being here.
1:07
This idea of thought leadership is really, really powerful. And I don't want
1:11
to get in the way of a very important speaker, Dr. Eric Horvitz.
1:15
But having him come and spend time with us and share his perspective is very, very powerful.
1:20
And he's a leading authority, obviously, on AI for decision-making.
1:25
He's also the chief scientific officer at Microsoft.
1:28
He spearheads the company-wide initiatives at Microsoft, navigating these opportunities
1:34
and challenges at the confluence of scientific frontiers, technology, and society,
1:40
especially including strategic efforts in AI, medicine, and the biosciences.
1:45
He's known for his contributions in AI theory and practice with a focus on principles
1:50
and applications of AI amidst the complexities of this marvelous open world that we live in.
1:56
And all of his efforts and collaborations have led to fielded systems in healthcare,
2:01
transportation, e-commerce, operating systems, aerospace, and I'm sure something
2:06
on my phone. I don't know what it is yet, but he'll tell us.
2:09
I also know that he's a member of PCAST, which means he's advising our government
2:14
about how to go forward in this space, which is also a very important thing.
2:18
Beyond his scientific work, Dr. Horvitz has pursued programs,
2:22
organizations, and studies on ethics, values, and safety with applications and influences of AI.
2:28
He founded and chairs Microsoft's Aether Committee on AI and looking at effects
2:34
and ethics in engineering and research.
2:36
He's also established the 100-year study on AI at Stanford and serves as a congressionally
2:42
appointed commissioner on the National Security Commission on AI.
2:46
So thank you for your public service, Eric, that's also powerful.
2:50
He's also received his PhD and MD degrees.
2:54
So I would call him a physician scientist actually from Stanford.
2:58
Notably, he's received the Feigenbaum prize and the Alan Newell prize for his
3:03
fundamental contributions to the science and practice of AI.
3:07
And I think you need to hear from him and not from me. So Eric,
3:11
come on up and thank you for joining us here in San Antonio.
3:14
And again, I want, I want to come back next year, please.
3:18
All right. Well, thanks so much for that wonderful introduction.
3:21
And it's wonderful to be here. I had to send my wife back a screenshot of the
3:26
weather forecast today.
3:27
At least last night, it said 95 in the Bay Area. She couldn't believe it.
3:32
I said, yeah, it's like summer here. It's fabulous.
3:36
Well, it's an honor to be here at the UT San Antonio Health Center and to share
3:40
some reflections on leveraging AI advances in healthcare in particular.
3:45
So in the next few minutes, I'll share some history, what I would refer to as
3:48
the arc of innovation with AI, with the particular lens on AI and medicine.
3:53
And then I'll touch on some key trends at the frontier that I think are very
3:57
interesting and informative for all of us in this field.
4:00
And then conclude with several opportunities and challenges ahead.
4:04
So first I see AI, I like to start with a definition, right?
4:07
You know, I like the definition that it's the study of computational mechanisms
4:11
underlying thought and intelligent behavior.
4:15
Now, in a document that kicked off the whole modern study of AI,
4:19
1955, the founders of the field said they were trying to find how to make machines
4:26
solve kinds of problems now reserved for humans.
4:30
Now, in 1959, just three years after the phrase AI was first used,
4:35
I think it was conjured up by John McCarthy, who I spoke to about that term.
4:39
But three years later, Robert Ledley and Lee Lustad wrote a foundational and
4:45
remarkably prescient article on the reasoning foundations of medical diagnosis in Science Magazine,
4:52
talking about how combinations of logic, probability, and decision theory could
4:57
aid our understanding of medical reasoning and the prospect of building computing
5:01
systems that could help physicians with diagnosis and therapy.
5:05
Now, that early focus on harnessing AI to enhance diagnostic accuracy is the
5:09
longest live and deepest aspiration for AI in medicine.
5:15
So it's now such an exciting time for applications of AI advances in medicine.
5:19
I just put up a little ontology.
5:20
Of course, we can have a bigger ontology in categories like this.
5:23
There are some great many efforts going on and a great many areas of opportunity,
5:27
from the use of AI in decision-making, to population health,
5:31
to biomedical research and science, Investigation to Education.
5:37
But that early focus on harnessing AI to enhance diagnostic accuracy is the
5:43
longest lived and deepest aspiration for AI in medicine.
5:47
So I'm going to focus my comments today on that holy grail of AI in medicine,
5:51
on directions with harnessing AI methods for boosting the quality and efficacy of clinical care.
5:57
So there's been a great deal of work in AI for medical decisions for over the
6:00
decades, but most work until the early 90s was on handcrafted systems.
6:06
In the late 1990s to today, a great deal of medical data became available,
6:11
in part because of the digitalization of medicine with electronic health records.
6:17
And that work, along with advances in machine learning algorithmically,
6:22
led to explorations of the use of machine learning to construct diagnostic and predictive models.
6:26
And to connect these models directly to the hospital information systems,
6:29
provide predictions based on fixed sets of evidence that were identified by
6:34
the ML machine learning algorithms from the data, and then used at runtime to make predictions.
6:39
So, for example, the health record might give a whole bunch of possibilities,
6:43
things to look at, tests and radiology imaging data and so on.
6:47
And the algorithms would pick a set of things to look at, given a goal prediction.
6:54
For example, in this case, one of our early pieces of work on predicting readmission
6:59
under 30 days, when that was coming to the fore as a penalizable metric.
7:05
It shows you the things that came up as a predictive evidence here.
7:08
You can run it in real time, and you can see how we did this study back in 2007, 2008.
7:14
We're looking at this paper on data-driven decisions for reducing readmissions.
7:19
Back in those days, when we submitted this to major journals in medicine,
7:21
people thought we were like, how does this work?
7:23
What are you talking about? What do you mean by machine learning?
7:26
Things have changed, though.
7:27
That system, that model underlied the readmissions manager, which we shipped
7:32
worldwide in 2009, I think. and it ran with its own test and train and test
7:37
algorithm at every site locally because we couldn't make it portable given different demographics.
7:42
But the probability of readmissions explained the information and so on.
7:46
It was based on data originally from the Washington Hospital Center,
7:49
a MedStar hospital in Washington, D.C.
7:52
Now, one thing about classical machine learning, and I just was talking with
7:55
Sam Beat about this at lunch,
7:56
that we still yearn for in the modern world of AI is these models could be well
8:02
calibrated, and they can give you well-characterized predictions with well-understood
8:06
sensitivities and specificities, which you need for any kind of medical device, for example.
8:11
So you can even operate at different thresholds here and understand what the
8:15
trade-off between true positive rate, for example, and false positive rate will be in a hospital.
8:19
We also did the same kind of modeling for the risk of C.
8:22
Diff, a hospital-associated infection, starting within 48 hours for patients
8:27
to prioritize, for example, antimicrobial stewardship.
8:31
The other really cool thing was when you had models that were well calibrated,
8:35
you can build this pipeline of data to predictions to decisions and do a decision-theoretic
8:40
cost-benefit model about allocating
8:42
resources based on the model's predictive power at any hospital.
8:47
Now, but 15 years ago, I mean, and this, the work I just talked about was ongoing
8:52
in the early 2000s. It continues to this day, and it's very promising.
8:56
But eyes have been turning elsewhere, given the flashiness and the excitement
9:00
about large-scale language models.
9:02
So about 15 years ago, we hit an inflection point in AI. And I think it's clear
9:06
that we've been riding on top of a fast-moving wave of innovation with deep
9:11
neural models since then.
9:14
And that wave has included reaching parity with humans on a number of challenging
9:18
benchmarks in visual and language tasks.
9:22
So around 2010, we started seeing the power of neural network models for applications
9:27
like image-based diagnosis in medicine, spanning dermatology and radiology,
9:32
running at expert levels for these tasks.
9:36
Where, for example, the AI could sort of highlight what it was looking at with
9:39
a particular heuristic and making its recommendations.
9:44
But importantly over the last four years
9:47
coming on five we've been riding on
9:50
yet another acceleration that i
9:53
see i'd like to talk about as being an
9:56
inflection upon an inflection with the qualitative and quantitative capabilities
10:01
we're seeing now with generative ai that are showing unexpected powers of generation
10:06
synthesis and composition the generative ai breakthrough have hinged on two powerful ideas.
10:14
Relatively straightforward, but they're clever. The first is the idea of self-supervision,
10:19
where massive corporate, like text sequences, represented by these squares here,
10:24
or imagery, can be harnessed for training via a hide-and-seek procedure,
10:28
where words or tokens, more generally, are randomly hidden by the system,
10:32
and the system's asked to guess them and gets better and better with a scoring
10:36
metric, capturing how often they can get the hidden tokens correct.
10:40
This lets you use the, instead of looking for an expert to label data for you,
10:44
you have the ground truth in front of you and you hide the ground truth and
10:48
train on massive corpora, much larger than we've trained on in the past.
10:53
The second innovation is teaching a model what to attend to via an architecture
10:59
named the transformer that is set up to allow models to learn what to look at across context,
11:05
some before and after, but typically before, and broader model activations as
11:09
it does a generation of the next tokens.
11:11
In this case, this system is trying to predict what comes after learning to
11:14
attend to context. What comes next?
11:18
Let's talk now about key trends. The two concepts I showed you are at the root
11:22
of key trends over the last five years, and these have great relevance to AI
11:27
for healthcare applications. We'll look at several today here.
11:29
First is the impressive power seen with increasing scale of compute and data.
11:35
There's been a lot of excitement over the last few years with what you see coming
11:39
out of these models that come with scale.
11:41
Here's the data center where GPT-4 was birthed, just west of Des Moines,
11:46
Iowa. and that data center had 10,000 GPUs and over 300,000 CPUs.
11:52
And at the time, it was one of the biggest GPU clusters in history.
11:57
Data centers like this are small now. We have much larger ones being built as
12:03
industry competes and pursues capabilities with scale.
12:08
So studies have demonstrated in some classic papers that with scale,
12:11
we see the rise of capabilities in language, mathematics, and programming tasks, for example,
12:16
with the sudden emergence of surprising abilities captured by the pops you see
12:20
here on this graph, like learning to solve a new task from a handful of examples
12:24
in a single interaction.
12:25
But more generally, the rise of what I think is the most magical capabilities
12:28
of abstraction, composition, and generalization.
12:32
So my team was one of the first in the world to examine GPT-4 carefully as part
12:38
of a request by OpenAI helped us understand this thing we just created.
12:42
My team was responsible for safety and capability reviews. And we were all shocked
12:47
in August of 2022 when we saw the general purpose powers of GPT-4 and the jump from GPT-3.5.
12:53
For example, showing the ability to write a proof of the infinitude of primes
12:56
as a poem, or to write computer programs in response to a prompt without specialized
13:01
training like you see here.
13:03
We're impressed with the ability to see effective responses to long,
13:06
detailed prompts, such as this one here, asking the model to perform a set of
13:09
complex visualization tasks.
13:12
Now, much is being done now with these and other large generative models,
13:15
but even deeper impact is on the horizon, and there's great application for
13:20
building workflows for examining, in this case, biomedical data sets.
13:24
But beyond the impressive powers that we saw with the generalization and abstraction
13:28
and code writing and poems, mathematical proofs inside of them,
13:33
we've been seeing powers of specialist expertise in the generalist models.
13:38
So we explored the power of the model for healthcare back in August of 2022,
13:42
in that same month. And for example, here is looking at some really challenging.
13:48
Diagnostic cases from the New England Journal of Medicine, the CPC cases here.
13:52
And we're amazed at the ability of these models out of the box to address some
13:56
of these most difficult challenges, as shown in this case of the diagnosis of babesiosis.
14:03
It's a tick-borne parasitic infection caused by Babesia macrotii for our parasitologists
14:08
in the audience that know about that.
14:10
We're also shocked by the ability of the model to solve medical competency exam questions.
14:14
We showed right away that there's a big jump before, including stellar performance
14:18
on step one, step two, step three of the USMLE exam.
14:21
I started to think about what I had to do to study for those exams back in the day.
14:27
Another interesting trend is the shifting of the Pareto frontier.
14:31
What I mean by that is with scale and other advances, we're seeing this trend
14:36
of this accuracy cost to compute frontier changing.
14:42
So x-axis here is the total cost. We're computing a result.
14:46
The y-axis is the accuracy. We're running on a very hard, big medical QA test set here.
14:53
And if you look from the right, lower right, and head up to the upper left,
14:58
you're seeing curves of what was possible at cost, bottom, and it's a log scale,
15:03
by the way, versus benefit accuracy.
15:06
Going from 2023 at the right to the middle, 2024, and now 2025.
15:12
So you see right now, for a few dollars, you can run this whole gigantic corpus
15:17
at higher accuracies than you could for $1,000 just a few years ago and with lower accuracy.
15:24
Another trend, now one of the hottest areas of research, is real-time or test-time reasoning.
15:31
It turns out that this was the focus of my PhD dissertation in the late 80s.
15:37
Controlling meta-reasoning for Bayesian networks of the day.
15:39
And I found myself sitting in meetings with OpenAI leads and at Microsoft talking
15:44
about taking principles from 1989 and 1990 and applying them to control how
15:49
we do runtime reasoning with these models with cost-benefits.
15:52
What you see on the left here is the power of the O1 model, one of the first
15:55
real-time reasoning models that does chain of thought.
15:59
And this is running on some challenging math problems as a function of training
16:03
time, building bigger and bigger models.
16:06
And on the right is what a smaller model does. You fix it and give it more and more test time compute.
16:12
So it's a whole new dimension opened up. This first came out,
16:16
it was internal and everyone was like so excited.
16:18
Like we can't tell, it's all a secret for a while.
16:21
You know, between opening Microsoft and we sort of rolled it out.
16:24
We had to see all the innovation happening around it. But it's not just graphs,
16:28
right? This is the power of real-time reasoning showing of medicine.
16:31
So I gave a little whisper to my friends at Harvard and Stanford,
16:35
my academic college, said, hey, let's, oh, one's coming out.
16:38
Let's get ready to, like, let's get ready. I just said, let's see how well these
16:42
runtime reasoning models do on medical hard cases here.
16:47
And so we wrote a paper that's, if you look up the exuberant title,
16:51
I wasn't a fan of the title, Superhuman Performance on the Reasoning of a Physician.
16:55
It's still under peer review, but it's an archive. You can read about these examples.
16:59
But what you see here is one example of the performance of O1 on a particular
17:04
heart medical challenge that's used in medical education called NEJM healer cases,
17:08
where O1's performance at the top here shows the performance of what's called the RRIDA score.
17:15
It's a psychometric score capturing the quality of clinical reasoning.
17:18
And GPT-4 is right beneath it. Just like, you know, one notch down in model
17:23
history, getting 47 out of 80.
17:26
And meanwhile, O1's getting 78 out of 80 on these very hard cases.
17:30
Another important trend, especially for AI methods in medicine, is multimodality.
17:35
So multimodal models can bring medical language and imagery together to build
17:40
joint representations.
17:41
And of course, radiologists, for example, are not siloed into imagery or language.
17:46
They look at both, and most people do on planet Earth.
17:50
We're going to weave these pillars into clinical medicine, into rich joint representations.
17:55
An early example was our work on our team on what's called LavaMed. Pretty amazing.
18:00
I mean, this took an open source model and then went out to PubMed and automatically
18:06
scraped tens of millions of figure-caption pairs, figure and caption,
18:10
tens of millions, and fine-tuned this model.
18:13
And all of a sudden, you can talk to the model about things it's seeing.
18:16
It's like we've given a visual cortex to the system by just post-training it
18:22
to understand these pairs of figure-caption.
18:25
Now, today we have access to multiple freely available models like Myra,
18:30
which is available on Azure, which provides visual grounding with its analysis.
18:36
It calls out automatically areas of attention as it talks about a radiograph,
18:41
for example, highlights them in different ways and calls things out.
18:46
Now, another exciting trend or sub-trend in multimodality is multifunctional imaging.
18:53
And I have to say, when I first saw the results of these models, it kind of blew my mind.
18:59
So what happens in the training phase for these models is they're combining
19:04
many biomedical imaging tasks, including across specialties.
19:11
The same model is learning about pathology, radiology, endoscopy.
19:15
It's also combining segmentation tasks, detection tasks, and recognition tasks.
19:21
It's almost like we're building a visual cortex of a radiologist,
19:24
but maybe one that was trained in residencies in many areas.
19:28
So as an example, this BiomedParse is built from 82 major biomedical imaging
19:36
tasks across nine imaging modalities from over 1 million images that are tagged
19:41
with segmentations and text labels.
19:43
As highlighted by the chart with different colors showing different areas of medicine.
19:48
So end users playing with this model, you can ask a large variety of questions to the same model.
19:52
In this case, on the left, histology and talking about where are the cancerous cells and so on.
19:59
On the right, brain imaging, same model.
20:02
We're asking it to highlight a region where there's a pathology.
20:06
And again, all in the same model.
20:09
Something I would not have expected to see, I don't know, even a few years after
20:14
the first BPG-4 ship, but certainly maybe not even in my lifetime.
20:18
Another trend is the power seen with generalized cross-modality models.
20:23
Notice I've changed from multimodal to cross-modality very deliberately.
20:31
And these extend beyond language-centric tasks and beyond vision,
20:35
where training and inference procedures bring together multiple layers of biomedical information.
20:40
So one example of this work is a project called Gigapath, which was published last year in Nature.
20:46
The project centered on training a multimodal model with over 150,000 pathology
20:50
slides, along with genomic data
20:53
linked to each slide, and cancer staging information linked to each slide.
20:58
And the model showed a scale how you can boost the performance on doing cancer
21:01
subtyping and mutation prediction just from the histology.
21:08
Now, over the longer term, I believe that these cross-modal models can provide
21:13
breakthrough approaches because they're weaving together different data sources,
21:18
figuring out how they go together in mysterious ways that no AI scientist,
21:23
I was talking with Samit about this,
21:24
no AI scientist understands today.
21:27
That's one of the differences between the Bayesian networks and the decision
21:30
theory I spoke about earlier.
21:33
And today we're working with models that we fundamentally are impressed by that
21:38
have some degree of characterizability, but we don't know how they work,
21:42
which is an interesting challenge to AI scientists.
21:47
Let's figure all this out because if we know how these things work,
21:50
we could probably do better with them.
21:51
But in this case, I'm saying that what I view as possible in our near future
21:55
is delivering ultra-personalized, ultra-diagnosis, therapy design,
22:03
and prediction of outcomes.
22:04
You're building models at population scales that combine imagery,
22:09
language, as well as the genomic, transcriptomic, proteomic,
22:13
and rich phenotypic information.
22:15
Talk now about another trend. People hear this phrase now a lot, gentic system.
22:21
We're seeing a rapid rise in what I would just say systems of models being used
22:26
where the models or model calls are interacting with one another,
22:30
provide a service or some rich set of capabilities.
22:34
We often refer to these multi-model calls or multi-agent systems as a gentic system.
22:40
We have frameworks freely downloadable from my team like Magentic One that allow
22:46
scientists and engineers to build more sophisticated problem solvers by defining,
22:50
what I would say, sets of specialists or agents to take on independent roles
22:55
and then orchestrating these agents into larger multi-step planners and solvers.
23:00
These agents act with different expertise, access to tools and data.
23:04
But we were just finishing up Magentic One and we didn't even have a name yet.
23:08
And I was going off to give a AAAS lecture called the OMEN lecture last year.
23:13
And I wanted to do a quick demo. So I worked with a couple of my engineering
23:15
friends. I said, this might be a kind of fun demo for a scientific group.
23:19
We called it the Biomedical Research Assistant, built on Pedantic One.
23:23
This is an example aimed at providing scientists someday as a prototype with
23:27
a biomedical research tool.
23:30
The system is powered by five separate agents, including a Web Surfer agent,
23:34
a bioscience research reviewer, and an orchestrator.
23:39
Talk about what these systems, these components all do.
23:42
Let me show you how this actually works. It's kind of fun to look at,
23:44
and it'll give you a sense for possibilities.
23:47
So the idea basically is to have an agent or the system collect and read relevant
23:52
literature, given a question or prompt, assess the scientific rigor of each article.
23:57
And I can talk about that. It's a whole agent in itself, you know,
24:00
double blinded, you know, how N equals what exactly, who funded it, and so on.
24:06
Synthesize a holistic report from all the papers, weighting the little scale
24:11
here, the veracity and the trust in the paper with the content,
24:16
go up and identify contradictions and disputes among these papers,
24:20
for example, and then propose forward steps with experiments and studies.
24:25
Say, for example, what is the role of telomere length in cellular aging and senescence?
24:30
What's the latest on that? Anybody know? I mean, it's unclear to me when I ask, question.
24:37
So let's read and collect relevant literature, begin to assess scientific rigor.
24:42
You see here a little snippet of the long report coming out on each article,
24:46
pull it all together and give me a holistic report weighing the veracity of
24:52
each article, identify contradictions and disputes.
24:56
Isn't that nice to know about? And then pose whole new steps of how can we resolve
25:01
these disputes? What's the next step on the frontier, for example?
25:05
So that was one of my first agent projects for my personal I was involved with.
25:09
But then in the last several months, we've gotten involved in several different agent-based projects.
25:15
One is on agents in diagnosis. We went back to the New England Journal of Medicine, these hard CPC cases.
25:20
We just published, right now it's under peer review, but we just published to
25:25
archive the results of this paper, the paper on these results.
25:30
And so what we did here is, again, looking at these hard New England Journal of Medicine CPC cases.
25:38
Typically, it's like this case, for example, a 12 pages long discussion with
25:44
a presentation, with lab tests, imaging studies, other information,
25:48
finally getting to sometimes a very rare diagnosis.
25:51
This was a 53-year-old woman with shortness of breath on exercise, kind of a mystery.
25:56
And the CPC cases are very interesting because if you are familiar with them,
26:00
you read them. And it's like they actually will bring together like a panel
26:04
of different physicians and independently and together.
26:06
And they all kind of scratch their heads and they think about the case and they
26:09
talk about what the findings are and so on.
26:12
So we employed a system we call the DX orchestrator, diagnostic orchestrator,
26:18
that employs five agents in a sequential diagnostic setting that work to get
26:23
to the final diagnosis while managing the costs of the tests.
26:28
So they're looking at minimizing costs, but getting to the answer,
26:32
efficaciousness, back to my earlier comments.
26:34
What are the five agents? Hypothesis agent. Maintains a probability rank differential diagnosis.
26:41
Test chooser agent. Selects tests that maximally discriminate among hypotheses.
26:46
Challenger agent. Continue to act as devil's advocate by identifying anchoring
26:52
bias, contradictory evidence, proposed tests that could rule out leading diagnosis, and so on.
26:58
Stewardship agent. Cost consciousness.
27:01
Advocates for cheaper alternatives. Sounds like an insurance company when diagnostically
27:06
equivalent tests are there for information value and it vetoes low yield expensive
27:10
tests like saying, come on, that's not going to tell us very much.
27:14
And a checklist agent with this quality control and it looks at test names coming
27:18
out so it knows how to collaborate with a database and so on.
27:22
All right, let's look at another Pareto frontier curve. At the bottom is the over 250 CPC.
27:30
These are the hard cases is the cost.
27:33
What was the mean total test cost? Get to the answer of the 250 cases.
27:39
And on the y-axis is the diagnostic accuracy reached across the 250 cases.
27:45
The blue shows you here the output for different configurations of the system.
27:51
Here you see on the curve, they see, for example, if we just gave the,
27:56
The DX orchestrator system, an $8,000 budget max, it's kind of operating at
28:02
about 80% accuracy and running about $2,500 of testing and so on.
28:07
The pink plus signs are the general practitioners we bought in to take the same test.
28:17
The plus sign is the mean on both the diagnostic accuracy on these hard tests
28:24
and the costs of tests being ordered by a human physician.
28:29
Now, we got a lot of blowback on this from some colleagues who said,
28:33
wait a minute, you didn't let the human beings have access to the web?
28:37
We said, nope, close book.
28:39
This is because these days, if you go to the web, you can't stop AI from creeping
28:44
in, right? Go to a Google search, there's the AI system.
28:48
But we decided to, in our next rev of the article, give folks access to some
28:56
of the standard medical tools that are used in practice.
29:01
And I have to say that, yes, I don't have the data here, but the plus sign does
29:05
move up a bit, but I think it moves to the right too.
29:09
It still is way dominated by this AI system here.
29:14
Now, on another agentic system for healthcare that our team is really deeply
29:18
immersed in right now, working with collaborators around the country, including Dr.
29:21
Kumar Shammar at UTSA, he's using this framework, from my understanding,
29:26
for looking at liver and renal diseases.
29:29
But we originally built the Healthcare Agent Orchestrator for helping with tumor board meetings.
29:35
For me, in medical school, I went back to after my PhD work, it took a few years,
29:42
I was blown away by the wonders of the collaborative or the collective intelligence
29:49
I saw at tumor board meetings.
29:50
So what happens at these meetings is very difficult cancer cases come to the
29:56
fore and a board of specialists,
30:00
all the relevant information, the pathologist, the surgeon, the radiation oncologist, the oncologist.
30:07
The social worker, the family, of course, the patient and the family will sit
30:11
together and you sit in a room.
30:14
In the old days, now, as we see here, it's Teams or Zoom.
30:18
But you sit in a room and a slideshow is presented radiology images showing
30:23
the cancer progression.
30:25
The histology, what are we seeing here? The stains, response to therapies and
30:30
so on. So it's complicated meetings right now.
30:32
This shows our collaborators at Stanford right now. Now, we started working
30:34
with Stanford originally on this, where literally we have IT,
30:38
we woven together, I say weaved together our IT groups.
30:43
So we kind of sit down and understand better how to build a better healthcare agent orchestrator.
30:48
So there's obviously a great deal of information across time with complex patients
30:52
with input from multiple specialties.
30:55
And we explored the use. The first use has been how the HAO system can be used
30:59
to pull together information to prepare clinicians for a productive tumor board meeting.
31:06
If an oncologist or resident is assigned to prepare a tumor board meeting, it's a big deal.
31:11
There's lots of information, history, notes to look through,
31:15
and lots of patient charts and so on.
31:17
And so what the system does here is you can actually call in different agents by name.
31:24
You know, patient history summarization agent, radiology agent,
31:28
pathology agent, staging agent, literature review agent, clinical trials agent, and so on.
31:33
For example, in this case, David Power, anonymized name here,
31:36
wrestling with prostate cancer, localized to seminal vesicles.
31:40
We bring the system in and we add, literally, especially with a command prompt,
31:46
add experts to the discussion from the system.
31:49
And then the system goes out and summarizes things that are per relevancy,
31:54
per its training and its prompting, to create supportive materials for helping with better decisions.
32:02
So the last topic I want to talk about today on trends resonates with the theme
32:08
of this meeting on augmenting human intelligence.
32:11
We have much to do in the realm to build AI systems that nurture and support
32:18
human decision-making and human agency, and that work to augment our intellect.
32:26
And we're still in the very earliest days of this work, unfortunately,
32:31
because it's such an important topic area.
32:33
I thought I would just tell a quick story about a recent study and its background here.
32:39
So in a recent study that I collaborated on as co-author based at Stanford,
32:44
we set up a randomized clinical trial that included 50 physicians in family
32:49
medicine, internal medicine, and emergency medicine at multiple sites,
32:53
including Stanford, Beth Israel,
32:55
Deaconess, Boston, and University of Virginia.
32:58
The clinicians were randomized into two groups, one using the LLM,
33:02
in addition to conventional diagnostic resources, the other only conventional
33:07
resources, books, search, and so on.
33:10
And we gave them six hard curated clinical vignettes, patient presentation, labs, and so on.
33:17
The finding is very intriguing here.
33:20
So the LLM did not significantly enhance the diagnostic reasoning performance
33:26
compared with the availability of only conventional resources here.
33:30
So conventional resource only, this is a JAMA network open publication here,
33:33
74%, using Gen AI, 76%, which I found very striking.
33:39
So I want to show this again more closely. He's like, show me that more closely, right?
33:45
So using Gen AI, the physician's got 76% on a rubric that goes to 100%.
33:51
Conventional only, 74%. Are you ready? Take a deep breath.
33:57
Gen AI alone. Gen AI alone.
34:00
This led to all sorts of crazy headlines in the New York Times.
34:04
AI chatbots, you know, it's like everyone got upset.
34:08
Well, the line that I wrote in the paper, among other lines,
34:11
that I said, let's get this line in there.
34:14
The L.M. Malone outperforming physicians really indicates that further development
34:19
in human-computer interaction is needed to realize the potential of AI in clinical
34:24
decision support system. You'd all agree, right?
34:27
So we went back to the July board. and working with Selene Everett,
34:31
Brian Bunning, two medical students, well, actually a PhD student and a medical
34:35
student for the last several months.
34:38
I said, let's leave the big team behind. Let's work as a threesome here.
34:44
So in a follow-on study, we were very careful to consider notions of complementarity
34:50
between humans and machines, a mix of initiatives, and working on grounding.
34:56
This was the seeking mutual understanding between human and machine.
34:59
Now, I've spent a number of years publishing on core papers in this space,
35:03
looking at the principles, decision theory, controlling who does what when with teams.
35:09
And we haven't really brought this to medicine yet.
35:12
So in our paper, which is also under review right now, hopefully be out soon,
35:16
on MedArchive called From Tool to Teammate, where the last study was on tool.
35:22
Now we're going to Teammate on a randomized controlled trial of clinician AI
35:26
collaborative workflows for diagnosis.
35:28
We set out in this way. We said, let's try different workflows.
35:32
Let's try AI as first opinion.
35:35
AI sees the vignette, generates its diagnoses as well as next best test to do
35:40
and final call on diagnosis.
35:43
The clinician looks at that, thinks about it, comes up with his or her own assessment.
35:49
Then we've asked the system to take both sources of information and weave,
35:56
weave together both inputs, display what came from whom.
36:03
Ah, the AI and clinician both agreed on polycythemia of Vera.
36:08
And also, as I said, comment, critique each one of these things here now.
36:12
Only the clinician thought EPO is a secreting tumor. And then the AI system
36:15
says, you know, but, you know, this kind of has a little argumentation going.
36:18
EPO makes this unlikely. However, you know, I guess rare analogs make it possible,
36:22
but less likely. And so on.
36:25
That's one workflow. The other is, by the way, you can imagine at some point,
36:30
the final diagnosis after the physician, the physician looks at this weave.
36:35
They now work with the patient, right? It's not the AI system.
36:40
Let's switch things around. Doctor goes first and then shares his or her findings with the AI.
36:47
AI goes second to the weave. So what we found now is going from the last paper to now,
36:56
the humans are doing different things, but on average, there's insignificant
37:02
difference between the AI and the human expertise using the system at this point.
37:07
And they're not copying the answers from the AI system. So it's interesting.
37:12
So there's much to do, and we're still analyzing the details of how the diagnostic
37:16
performance of clinicians using conventional resources, which is the middle column here.
37:20
This is like what you see here is six cases of the vignettes.
37:24
The middle column shows the scores that clinicians came up with,
37:28
how many of them out of the full battery came up with like a top score versus
37:34
second, top, third, top.
37:36
These were actionable scores, a portion of the scores. but the middle is the
37:41
conventional and then we have the second opinion doctors acting after AI on
37:47
the right hand column here.
37:48
So you see interestingly some of the low scores in the Sankey plot are going
37:53
up after exposure to AI but interestingly some are coming down.
37:57
Isn't that interesting? So what we're trying to do here given the theme of this
38:02
gathering today on augmentation there is a rich.
38:06
Structural like fabric between machines and humans that design principles and
38:14
study can help us better understand,
38:17
so with that let me just say that.
38:20
We have many challenges and opportunities ahead. I look at like the thing,
38:23
small C, big O, you know, being reasonable.
38:28
I would say that at the top of the list for me is human-eat collaboration. Let's figure that out.
38:33
Let's think about human agency. And by the way, figuring that out,
38:38
I think, is not just technical. It's socio-technical.
38:41
Like, how is medicine going to evolve? What does it mean to be a physician?
38:46
What residency program do medical students select and why? What do they expect
38:51
they'll be doing in those roles?
38:53
We need, per my conversation with Sam bit earlier, we need to figure out how
38:57
to, on this new dark matter called deep neural network, to figure out how to
39:02
make them reliable, how to trust them, and how to evaluate them well.
39:08
Then, thinking about, like, it's never the case that you have a box sitting
39:12
around coming out of UTSA computer science lab that's going to be delivered
39:15
to a doctor to the hospital and have it be used or trusted.
39:19
We have to think about workflow integration and, spelled more correctly here,
39:24
a diffusion model of how you diffuse these technologies into people,
39:27
with people, society, and organizations.
39:29
And we have a lot of work to do on norms, ethics, and regulation.
39:32
I was involved in this interesting National Academy of Medicine report called AI Code of Conduct.
39:37
I'd love you to take a look at that. We try to make it at the right level of
39:40
abstraction to make sense and not be overwrought with regulatory kind of suggestion.
39:47
So I want to end by calling out attention to a set of reflections captured in
39:53
an edited piece that I put together named the AI Anthology a couple of years ago.
40:00
We did this AI Anthology before GPT-4 was public in November of 2022.
40:08
I said, oh my God, I can't sit on this, this model. Can I get permission?
40:12
I know it's all top secret.
40:13
Can I get permission to share the model out with experts across fields?
40:17
Just get their reactions to this.
40:20
So you'll see it's a pretty broad span of experts from across the world,
40:24
including across the sectors, Marshall McNutt, National Academy of Science,
40:29
the dean of engineering at Duke,
40:32
psychologists, psychiatrists, mathematicians, physicists.
40:37
But I want to say that there's four folks I asked from the medical area.
40:43
I gave them this piece of magic to look at and asked them to write a two to
40:48
three page essay, something a four page essay.
40:50
Just what I asked them to do was give us your reactions to what you're seeing.
40:57
And please, how can we write? How can we harness this for the benefit of humankind?
41:03
What are the best? What do we need to do? And so these four folks here.
41:09
Natalie, Trevor, Yaa, and Robert commented.
41:12
I want to just end with Bob Wachter's comments. He was then chief of medicine at UCSF.
41:18
He titled his essay, Will GPT-4, but you can put in a little large language
41:22
model, be the technology that finally transforms healthcare for the better?
41:27
Will future versions of GPT-4 be the killer app that finally allows healthcare
41:32
to deliver new levels of quality, safety, and equity, and improve patient and provider experiences?
41:38
That's how he ended. And will it generate efficiencies that allow our healthcare
41:42
system to operate at a cost that doesn't bankrupt businesses,
41:45
government, and patients?
41:47
I believe the answer will be yes. But it exceeds the capacity of this particular
41:52
human to predict whether these benefits will be seen in two years or 20.
41:57
So I think we could all come together to bring many of these possibilities for
42:02
enhanced healthcare sooner rather than later.
42:06
Thank you. Thanks for listening to that fabulous talk.
42:11
Don't forget to click on the link for free credit.
42:14
I'm Holly Wehment. I'll see you soon.
42:17
Pediatrics Now is brought to you by the Department of Pediatrics at UT Health
42:21
San Antonio and University Health's new Women's and Children's Hospital.
42:27
Our website is pediatricsnowpodcast.com.